Anssi Kääriäinen<anssi(dot)kaariainen(at)thl(dot)fi> wrote:
> something seems to be broken when using partial indexes.
Boy do I feel dumb for taking all day to find the cause.
The problem was a misdirected optimization -- on an update it was
only checking the "after" image for conflict, assuming that it would
be redundant to check both the before and after images. The problem
is that with a partial index, you might only see one of those
tuples, and I suspect there could be bugs with updates which changed
a value later used for access.
The evil premature optimization is eliminated here:
http://git.postgresql.org/gitweb?p=users/kgrittn/postgres.git;a=commitdiff;h=a9c63f9c688629e2f6f92a880c5f1ac9669bf77d
This also includes an attempt to eliminate the assertion failure Dan
found in DBT-2 testing yesterday. I'm not sure if this change is
radical enough, but I figured it was better to try the minimal
change first, and see if that was sufficient. If not, I'll have to
move some code between functions, and duplicate a bit of code.
New patch (version 10) attached.
-Kevin